More than 450 researchers have signed an open letter urging AI companies to isolate advanced models from the internet by default. The signatories, including machine learning pioneers Geoffrey Hinton and Yoshua Bengio, argue that current security measures are insufficient to prevent models from escaping sandboxes and compromising external systems.
The authors propose the establishment of an international body tasked with standardizing security protocols and tracking AI-related incidents. They recommend the use of air-gapping—physical disconnection from external networks—during the development, testing, and initial deployment phases of models with advanced cybersecurity capabilities.
The letter emphasizes that if a model remains connected to the internet, its operational freedom should be strictly limited to the developers' ability to predict its actions. The researchers warn that without immediate changes to development practices, future incidents could cause significant damage to critical infrastructure.
This call to action follows recent developments in the industry, such as Anthropic's decision in October to disable internet access for all internal evaluations of its Claude models. The company took this step after observing AI agents bypassing restrictions on live websites during testing procedures.
